Query-Side BM25 Retrieval Model

• Query-Side BM25 is a symmetric adaptation of BM25 that weights both query and document tokens to normalize term frequency and boost retrieval accuracy.
• It reduces overrepresentation of repeated or common tokens by applying BM25 calculations to query terms, enhancing the discriminative power of key terms.
• Empirical results on benchmarks like BRIGHT demonstrate improved nDCG scores and relevance in retrieval-augmented generation scenarios.

Query-side BM25 refers to the adaptation of the BM25 retrieval model in which the BM25 scoring function, traditionally applied to documents, is also applied to the query—yielding a fully symmetric, BM25-weighted query representation. Rather than using a raw bag-of-words (token frequency) vector for the query, each query token’s weight is calculated via the BM25 formula based on its frequency and normalization, matching the weighting applied to document tokens. This approach is motivated by the observation that as queries become longer and more reasoning-intensive—such as in the BRIGHT benchmark or in large prompts driving retrieval-augmented generation—bag-of-words weighting overrepresents repeated or uninformative tokens, diminishing retrieval accuracy. Query-side BM25 normalizes the influence of each query term and improves the fidelity of document ranking, primarily in settings with long or complex queries (Ge et al., 2 Sep 2025).

1. Principles and Motivation

Classic BM25, as implemented in systems like Anserini and Pyserini, uses the BM25 formula exclusively on the document side: each query token is assigned a raw frequency in the query vector, while the document vector reflects BM25’s probabilistic frequency, inverse document frequency (IDF), and document length normalization. When queries are short, the distinction is inconsequential: weighting is heavily dominated by IDF and low-frequency tokens. However, as query length increases, repeated or generic tokens in the query can dominate relevance scores if left unnormalized.

Query-side BM25 (“query BM25” or “symmetric BM25” [Editor’s term]) mitigates this by assigning each query token a BM25-calculated weight, suppressing the outsized influence of repeated/uninformative tokens and amplifying the contribution of meaningful, rare tokens. This change is particularly impactful in modern IR where queries often include extensive, naturalistic language (e.g., LLM prompts or reasoning tasks), far exceeding the 2–4 keyword queries for which classical retrieval models were originally calibrated (Ge et al., 2 Sep 2025).

2. Formal Scoring Algorithm and Implementation

Let $t$ be a token and $x$ a text (document or query). The BM25 term weight is:

$$\text{weight}(t, x) = \mathrm{IDF}(t) \cdot \frac{tf(t, x) \cdot (k_1 + 1)}{tf(t, x) + k_1 \cdot (1 - b + b \cdot \frac{\text{len}(x)}{\text{avg\_len}})}$$

In classical BM25, this is computed for document $d$ only, and the query vector $q$ is $\text{tf}(t, q)$ (raw count). In query-side BM25, the same formula is applied to $q$ as well, producing a BM25-weighted query vector, so final scores are computed as:

$$\text{score}(q, d) = \sum_{t \in V} \left( \text{weight}(t, q) \cdot \text{weight}(t, d) \right)$$

where $V$ is the vocabulary. In practice, implementations expose this as a switchable option (e.g., via BM25QueryGenerator in Anserini, or via Pyserini’s BM25 + Gensim integration) (Ge et al., 2 Sep 2025).

Notably, because both query and document are subject to identical normalization and term-frequency nonlinearity, the scoring is strictly symmetric. Unlike bag-of-words, where repeated terms linearly increase a token’s query weight, query-side BM25’s normalization term diminishes marginal returns.

3. Comparison to Traditional BM25 and Performance

Traditional BM25 treats only document terms with the full weighting, using raw frequency vectors for queries. For short keyword queries, this produces similar results as the BM25 effect saturates quickly. However, for medium-to-long queries (16–256 tokens with repeated or context-heavy phrasing), traditional BM25 overweighs common terms and fails to discriminate key reasoning tokens—a problem increasingly prevalent in reasoning-intensive benchmarks and retrieval-augmented generation.

Empirical results (Ge et al., 2 Sep 2025) show that, for these longer and more complex queries:

• Query-side BM25 provides significantly better normalization of token influence,
• nDCG@10 and related metrics frequently improve relative to bag-of-words weighting,
• The largest gains (in both win frequency and score delta) are seen in medium-length queries, with diminishing effects for extremely short queries or extremely long ones.

For short queries, both approaches converge; for queries with highly variable term frequencies—as is common in user-generated or LLM-generated prompts—query-side BM25 more faithfully reflects term discriminativeness.

4. Impact on Retrieval Pipelines and Alignment with LLM-Augmented Scenarios

The motivation for query-side BM25 is further strengthened with the adoption of long, natural language queries in retrieval-augmented generation (RAG) pipelines. LLM-generated prompts can far exceed classical query lengths, and simple bag-of-words representations lead to degraded performance as observed in the BRIGHT benchmark (Ge et al., 2 Sep 2025). By assigning length-normalized and IDF-weighted influence to each query token, query-side BM25 produces more discriminative candidate rankings.

When integrating with reranking architectures such as RankLLM or listwise LLM reranking:

• Initial retrieval sets generated with query-side BM25 contain more relevant candidates,
• Listwise rerankers—conditioned on higher-quality input sets—yield substantially improved nDCG@10 and other metrics post-reranking,
• The improved alignment between retrieved sets and reasoning-intensive queries enables LLMs to generate more contextually appropriate responses in RAG settings.

This approach produces retrieval results closer to the metrics seen in BRIGHT’s own evaluation, which explicitly uses query-side BM25 for its baselines.

5. Practical Considerations, Limitations, and Parameter Tuning

Enabling query-side BM25 requires modifications to query generation and weight calculation in retrieval toolkits. In Anserini and Pyserini, dedicated query generator classes support this with consistent parameterization (the same $k_1$ and $b$ as for documents). Caution should be taken to maintain consistent document length normalization (“accurate” vs. “quantized”) between queries and documents, as mismatches may lead to suboptimal weighting and degraded performance (Ge et al., 2 Sep 2025).

For downstream consumers (e.g., Pyserini’s interface with Gensim or RankLLM), the main requirement is to preserve tokenization and token statistics between index construction and query evaluation.

A current limitation is that most research has focused on English and the BRIGHT dataset; validation across BEIR and other benchmarks is needed to broadly quantify generalization. Additionally, query-side BM25 does not introduce semantic modeling; its gains are primarily derived from improved lexical weighting in long queries.

6. Prospective Directions and Research Implications

The adoption of query-side BM25 suggests multiple future research trajectories:

• Revisiting foundational IR paradigms to view both query and document under identical probabilistic weighting,
• Investigating the role of symmetric normalization for complex or agentic multi-turn queries that arise in LLM-powered augmentation,
• Exploring hybrid pipelines where query-side BM25 is composed with advanced neural reranking, query reformulation, or expansion techniques,
• Assessing efficient and hardware-optimized implementations in both batch and streaming architectures, as longer queries and prompt-based retrieval become standard.

A plausible implication is that as retrieval-augmented generation expands and queries become increasingly long and structured, query-side BM25 (or similar symmetric models) will become an essential baseline for robust, interpretable, and reproducible IR pipelines.

Table: Key Distinctions—Traditional vs. Query-Side BM25

Aspect	Traditional BM25	Query-Side BM25
Query Vector	Bag-of-words (raw counts)	BM25 scoring per token
Term Weight	Linear in query freq	Length/IDF-normalized
Long Query Handling	Overweights repeated tokens	Suppresses redundant terms
Scoring Symmetry	Asymmetric	Symmetric
Best for	Short queries (keywords)	Long/complex queries

Conclusion

Query-side BM25 enhances the classic retrieval model by BM25-weighting both query and document tokens, normalizing term contributions for long or repetitive queries, and enabling superior ranking quality particularly for reasoning-intensive and LLM-derived contexts (Ge et al., 2 Sep 2025). Its integration into major toolkits (Anserini, Pyserini) and its empirical validation on the BRIGHT benchmark underscore its practical significance for next-generation information retrieval systems.